The biggest upsets in our data, and why they aren't surprising
By FootInsights · Published · Updated 26 Sep 2026 · 5 min read
Every season produces a result that seems to break the sport: a champion beaten at home by a side fighting relegation, at a price nobody would have taken seriously. Those matches get remembered as evidence that football is chaos and prediction is futile.
Our own data says something more interesting. The upsets are real, they are memorable — and they arrive at close to the rate the market said they would.
The longest-priced winners we hold
Every finished match in our database carries a closing price for each outcome. Ranking them by the price of the outcome that actually happened, these are the biggest shocks across the nine leagues we cover:
| Match | Result | Date | Winner’s closing odds | Implied |
|---|---|---|---|---|
| Bayern Munich – Werder Bremen | 0–1 | 21 Jan 2024 | 26.00 | 3.8% |
| Bayern Munich – Eintracht Frankfurt | 1–2 | 3 Oct 2021 | 24.82 | 4.0% |
| Barcelona – Las Palmas | 1–2 | 30 Nov 2024 | 23.00 | 4.3% |
| Barcelona – Leganés | 0–1 | 15 Dec 2024 | 22.93 | 4.4% |
| Manchester City – Crystal Palace | 0–2 | 30 Oct 2021 | 21.00 | 4.8% |
| Manchester City – Brentford | 1–2 | 12 Nov 2022 | 19.50 | 5.1% |
| PSV – Telstar | 0–2 | 30 Aug 2025 | 19.35 | 5.2% |
| Benfica – Portimonense | 0–1 | 3 Oct 2021 | 18.19 | 5.5% |
Every one is an away win at a dominant club’s home ground, which is the only place these prices exist. You need one team the market considers overwhelming and another it barely rates, and European football’s biggest clubs playing at home is where that gap opens widest.
Notice the implied probabilities, though. The very biggest shock in five seasons across nine leagues was priced at 3.8% — roughly a one-in-26 chance. Not a miracle. A one-in-26 chance, in a dataset containing thousands of matches where one side was a heavy favourite.
Upsets are the schedule, not the exception
Here is the part that changes how the table above should be read. Grouping every outcome in our data — home, draw and away alike — by its closing price:
| Closing odds | Outcomes | Actually won | Price implied |
|---|---|---|---|
| Under 5.00 | 41,727 | 37.38% | 38.37% |
| 5.00 – 7.99 | 5,501 | 15.52% | 16.74% |
| 8.00 – 11.99 | 1,704 | 7.81% | 10.67% |
| 12.00 – 19.99 | 748 | 4.95% | 6.92% |
| 20.00 and above | 204 | 2.45% | 4.07% |
Longshots win. Outcomes priced between 8.00 and 12.00 came in nearly 8% of the time — about one in thirteen. Across a season of fixtures that is a steady drumbeat of “shock” results, each individually improbable and collectively entirely expected.
What feels like chaos is mostly availability bias: you remember Bayern losing at home to Werder Bremen and you do not remember the several hundred matches where the heavy favourite duly won.
What the gap between the columns is
The two right-hand columns never match, and the difference is not the market being wrong. A bookmaker’s implied probability includes their margin, so it sits above the true probability by design — that gap is the fee, and stripping it out is the first step in reading any price honestly.
The revealing part is how unevenly that fee is spread. Expressed as a share of the implied probability, the overcharge runs:
- Under 5.00 — about 3% above true
- 5.00 – 7.99 — about 7%
- 8.00 – 11.99 — about 27%
- 12.00 – 19.99 — about 28%
- 20.00 and above — about 40%
That ordering is the favourite-longshot bias, and this is it measured in real closing prices rather than argued from a hypothetical. Back a short price and you are paying a small fee. Back a 20.00 shot and you are paying something closer to two-fifths of the outcome’s actual worth.
Two cautions on that last row: it holds only 204 outcomes, and 2.45% of 204 is five winners. Small samples move a lot, and that is not a number to lean on. The direction is consistent across all five bands, which is the finding; the exact figure at the extreme is not.
What this means in practice
A model being “wrong” about an upset usually isn’t an error. If a fixture is priced at 4% and the 4% happens, nothing failed — the estimate was about how often, not about which one. This is why accuracy is the wrong metric for probabilistic forecasts, and why our track record scores calibration rather than counting hits.
Longshot value is where value looks most attractive and is least likely to be real. The biggest apparent edges cluster at long prices because that is where the fee is largest and where a model’s own estimates are least reliable. That combination is why most apparent edges are model error.
The market is well calibrated where it matters. At prices under 5.00 — nearly nine-tenths of all outcomes — the market’s estimate sits within about a point of what actually happens once its fee is accounted for. That is a difficult standard, and it is the one any prediction service should be measured against.
About these numbers
Every figure comes from our own database: finished matches in the nine leagues we cover, joined to closing prices from football-data.co.uk. It is a snapshot as this was written and will drift as more matches finish. The track record grades every prediction we publish against reality, misses included.
Upsets are part of the distribution rather than a break in it. Treat probabilities as probabilities, and stake only what you can afford to lose.